In this paper, we propose a new defect detection method using a deep autoencoder
and one-class support vector machine. The proposed method extracts
patches in insert images and classi_x000C_es each patch into normal and defect one.
However, the appearance of defects varies from case to case and it is very di_x000E_cult
to collect all possible defect patch images, which hinders the use of conventional
binary classi_x000C_cation methods. Therefore, we develop a novel method that only
requires normal patches. To be precise, the method uses a deep auto-encoder as a feature extractor, which is trained with only normal patches, and one-class SVM
is adopted to determine the decision boundary of normal cases. Experimental
results show that the proposed method works robustly for light changes and
improves the classi_x000C_cation performance compared with conventional methods.
